GPTomics/bio-admet-prediction

Version Compatibility

Reference examples tested with: RDKit 2024.09+, requests 2.31+, DeepChem 2.8+, chemprop 2.0+ (note major API change from 1.x), admet-ai 1.3+, pandas 2.2+.

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

ADMET Prediction

Predict absorption, distribution, metabolism, excretion, and toxicity properties of drug candidates. ADMET prediction underpins lead selection and de-risking; calibrated, applicability-domain-aware predictions distinguish a working filter from a costly false-confidence rejection. Modern best practice combines online services (ADMETlab 3.0 with uncertainty estimates), open-source models (chemprop D-MPNN), and rule-based filters (Lipinski / Veber / BBB heuristics) -- each with known failure modes.

For PAINS / Brenk / structural alerts, see chemoinformatics/substructure-search. For QSAR model building from in-house data, see chemoinformatics/qsar-modeling.

ADMET Model Taxonomy

| Tool | Endpoints | Architecture | Uncertainty | Access | Fails when | |------|-----------|--------------|-------------|--------|------------| | ADMETlab 3.0 | 119 reported features: 77 prediction models, 34 computed properties, 8 rules | Multi-task DMPNN + descriptors for modeled endpoints | Evidential uncertainty for modeled endpoints | Web service; hosted API documented by the authors | Outside training distribution; metals; macrocycles | | ADMET-AI | TDC-derived ADMET tasks; inspect installed model metadata | Chemprop D-MPNN | Inspect version-specific outputs; do not assume calibrated uncertainty | Python package | v2 package predictions differ from the v1 paper/server | | DeepChem MolNet | Dataset-dependent tasks including Tox21, ToxCast, and ClinTox | Model-dependent | Model-dependent | Python package | Coverage and uncertainty depend on the selected dataset/model | | pkCSM | Service-defined ADMET endpoints | Graph signatures + ML | Inspect current service output | Web service | Applicability domain and service contract must be checked | | SwissADME | Physchem, pharmacokinetics, drug-likeness, and medchem outputs | Published models and rules | None advertised | Web service (no public API) | Automated access is restricted by its terms | | ProTox-3.0 | 61 toxicity models/endpoints | RF/DNN + fingerprints, similarity, and pharmacophore methods | Confidence score | Web service / sample API | Toxicity only; reports LD50 and toxicity class | | ADMETpredictor (Simulations Plus) | ~140 | Proprietary | Per-prediction | Commercial | License cost | | FAF-Drugs4 | filters | Rule-based | None | Web | Static rules | | chemprop (in-house) | User-defined | D-MPNN ± descriptors | Ensemble and other estimators; optional calibration | Python package | Requires suitable training and calibration data |

Decision: For batch screening with no in-house data, ADMETlab 3.0 provides 119 reported platform features and uncertainty for modeled endpoints through its web service and hosted API; verify the live API documentation before automating access. For a sufficiently large, relevant in-house endpoint dataset, benchmark a chemprop D-MPNN, descriptors, and simpler baselines under a deployment-relevant split rather than assuming a universal sample-size threshold. Shan et al. (2022) reported an AUC of 0.956 for a D-MPNN combined with 206 MOE descriptors on their random-split hERG benchmark.

Decision Tree by Scenario

| Scenario | Workflow | Reasoning | |----------|----------|-----------| | Library triage, no in-house data | ADMETlab 3.0 API batch | Broad platform coverage plus modeled-endpoint uncertainty | | Single endpoint, adequate in-house data | Benchmark chemprop D-MPNN, descriptors, and simpler baselines | Select by prospective or deployment-relevant validation | | Need calibrated probabilities | chemprop with ensemble + Platt | Native deep learning rarely calibrated | | FDA / regulatory submission | OECD-compliant QSAR with AD | See OECD principles below | | Quick annotation for VS | Lipinski, Veber, and QED reported separately | Rank or annotate; do not impose a universal QED gate | | BBB penetration | Simple screen: TPSA <= 90, MW <= 500, HBD <= 3 | Repository heuristic; not the six-factor CNS MPO | | Cardiotox liability | ADMETlab hERG + ProTox-3.0 cardiotoxicity + literature check | Both hosted endpoints model hERG blockade; compare applicability domains and assay definitions | | Drug-drug interaction (CYP) | CYP1A2/2C9/2C19/2D6/3A4 inhibitor + substrate | Standard set of 5 CYPs |

OECD QSAR Principles (5 Pillars)

For regulatory-grade ADMET QSAR (REACH, ECHA, FDA submissions), models must satisfy:

1. Defined endpoint -- specific bioassay, units, conditions
2. Unambiguous algorithm -- reproducible model + code
3. Defined applicability domain (AD) -- where the model is valid
4. Appropriate statistical validation -- external test set, cross-validation
5. Mechanistic interpretation -- biological / chemical rationale

For non-regulatory work, AD assessment is still critical. The OECD's applicability domain is the workhorse: predictions outside the AD are unreliable, but operational AD measures (leverage, kNN, conformal prediction) often disagree.

Applicability Domain Methods

| Method | Definition | Flags out-of-AD when | |--------|-----------|----------------------| | kNN distance | Mean distance to k nearest neighbors in training set | > training-set distribution P95 | | Leverage (Williams) | Hat-matrix diagonal | > 3p/n (p = features, n = compounds) | | Density (KDE on PCA) | Density in feature space | < density of training set P5 | | Conformal prediction | Per-prediction confidence interval | Interval > tolerance | | Bayesian variance | Ensemble or MC-dropout variance | > training-set variance P95 |

For deep-learning ADMET, conformal prediction can provide calibrated prediction sets or intervals when its exchangeability and calibration assumptions are appropriate (McShane et al. 2024).

ADMETlab 3.0 API

ADMETlab 3.0 reports 119 platform features: 77 prediction models, 34 computed physicochemical properties, and 8 medicinal-chemistry rules. The modeled endpoints include prediction uncertainty; do not imply that computed properties and rules have model uncertainty.

Goal: Obtain the platform's 119 features for a batch of SMILES, including uncertainty for the 77 modeled endpoints, using the hosted service.

Approach: Follow the live ADMETlab 3.0 API tutorial to wash molecules, submit batch predictions, and retrieve the returned results. The 2024 paper documents API/batch support and modeled-endpoint uncertainty. Obtain current rate limits, routes, payloads, task identifiers, and output contracts from the live official documentation rather than attributing them to the paper or hard-coding an unofficial example.

import pandas as pd

# After submitting with the current official API example, load its CSV output.
results = pd.read_csv('admetlab3_results.csv')
# Preserve the uncertainty columns and task identifier in downstream reports.

ADMETlab 3.0 endpoints (sample):

• Absorption: Caco-2 permeability (logPapp), HIA (%), Pgp inhibitor/substrate, MDCK
• Distribution: BBB+, PPB (%), VDss (L/kg), Fu (fraction unbound)
• Metabolism: CYP1A2/2C9/2C19/2D6/3A4 inhibitor / substrate
• Excretion: CL (mL/min/kg), T1/2 (h)
• Toxicity: hERG, AMES, hepatotoxicity (DILI), carcinogenicity, immunotoxicity, mutagenicity, respiratory, skin, eye, cardiotoxicity, mitochondrial, NR-AR, NR-ER, SR-MMP
• Drug-likeness: Lipinski, Veber, Ghose, Egan, Muegge, QED, SAscore

chemprop D-MPNN for Custom Endpoints

When in-house data is available, train a target-specific model. chemprop provides a widely used open-source D-MPNN architecture with atom/bond features and optional molecular descriptors; benchmark it against appropriate baselines on the project's data.

Goal: Train a target-specific ADMET classifier or regressor on in-house bioassay data.

Approach: Use the installed Chemprop 2.x CLI with a scaffold split, a release-supported descriptor featurizer, replicated models, and an explicit prediction-time uncertainty/calibration workflow.

# Chemprop 2.2 CLI; verify flags against the installed release.
# chemprop train --data-path data.csv --task-type classification \
#                --save-dir model_dir --split-type scaffold_balanced \
#                --molecule-featurizers v1_rdkit_2d_normalized \
#                --num-replicates 5 --ensemble-size 5
# chemprop predict --test-path test.csv --model-paths model_dir \
#                  --uncertainty-method ensemble --preds-path predictions.csv

# Or chemprop 2.x programmatic API (full programmatic API documented at chemprop.readthedocs.io)
# See chemoinformatics/qsar-modeling for the full chemprop 2.x training pipeline.

Key: Replicates and an ensemble estimator produce an uncertainty estimate, not automatic calibration. Fit and evaluate a documented calibrator on a separate calibration set when calibrated probabilities or intervals are required. Descriptor benefit must be demonstrated on the intended endpoint.

hERG Cardiotoxicity Endpoint

hERG (KCNH2) blockade can contribute to QT prolongation and Torsades de Pointes and is an important non-clinical cardiac-safety endpoint. Follow the current ICH S7B/E14 and regulator-specific guidance applicable to the program rather than treating one model output as a regulatory conclusion.

| Model | Architecture | Training data | AUC | Reference | |-------|--------------|---------------|-----|-----------| | Shan et al. D-MPNN + MOE | D-MPNN + 206 MOE descriptors | 7,889 compounds | 0.956 (random split) | Shan 2022 | | CardioTox-net | Five DL base representations + neural meta-ensemble | BindingDB, ChEMBL, and literature | 0.930 (10-fold meta-validation) | Karim 2021 | | ADMETlab 3.0 hERG | DMPNN multi-task | Internal | 0.92 (reported) | Fu 2024 | | ProTox-3.0 cardiotoxicity | RF-based classifier | 5,252 ChEMBL compounds with hERG IC50/Ki | 0.86 CV; 0.95 external | Banerjee 2024 |

Interpretation: A single-model probability > 0.5 is NOT a kill signal. Triangulate multiple hERG-specific models and a literature search. ProTox-3.0 calls the endpoint cardiotoxicity, but its model specifically predicts small-molecule hERG blockers; it should not be treated as an independent non-hERG mechanism. Consider exposure relative to measured hERG potency and confirm important decisions experimentally rather than applying a universal safe/unsafe IC50 cutoff.

CYP Inhibition (DDI Risk)

5 CYP isoforms cover most clinically relevant DDIs:

| CYP | Substrates (drugs) | Inhibitor flag if predicted prob | Action | |-----|--------------------|-----------------------------------|--------| | CYP3A4 | many drug classes | Model-specific threshold | Interpret inhibitor and substrate assays separately | | CYP2D6 | beta-blockers, antidepressants | Model-specific threshold | Include polymorphism and exposure context | | CYP2C9 | warfarin, NSAIDs | Model-specific threshold | Evaluate clinical substrate/exposure context | | CYP2C19 | PPIs, clopidogrel | Model-specific threshold | Include polymorphism and assay context | | CYP1A2 | caffeine, theophylline | Model-specific threshold | Include induction, diet, and smoking context |

PAINS, BRENK, REOS Filters

ADMET prediction is separate from structural alerts; combine. See chemoinformatics/substructure-search for PAINS/BRENK/REOS pattern catalogs.

from rdkit.Chem.FilterCatalog import FilterCatalog, FilterCatalogParams

def alerts(mol, catalogs=('PAINS_A', 'BRENK', 'ZINC')):
    params = FilterCatalogParams()
    for cat in catalogs:
        params.AddCatalog(getattr(FilterCatalogParams.FilterCatalogs, cat))
    catalog = FilterCatalog(params)
    hits = catalog.GetMatches(mol)
    return [h.GetDescription() for h in hits]

Lipinski / Veber / Drug-Likeness

See chemoinformatics/molecular-descriptors for full physchem table. Quick filter:

from rdkit.Chem import Descriptors, Lipinski, QED

def druglike_score(mol):
    mw = Descriptors.MolWt(mol)
    logp = Descriptors.MolLogP(mol)
    hbd = Lipinski.NumHDonors(mol)
    hba = Lipinski.NumHAcceptors(mol)
    tpsa = Descriptors.TPSA(mol)
    rotbonds = Lipinski.NumRotatableBonds(mol)
    qed = QED.qed(mol)

    lipinski_violations = sum([mw > 500, logp > 5, hbd > 5, hba > 10])
    veber_pass = rotbonds <= 10 and tpsa <= 140
    bbb_simple_screen = tpsa <= 90 and mw <= 500 and hbd <= 3

    return {'MW': mw, 'LogP': logp, 'HBD': hbd, 'HBA': hba,
            'TPSA': tpsa, 'RotBonds': rotbonds, 'QED': qed,
            'Lipinski_violations': lipinski_violations,
            'Veber_pass': veber_pass, 'BBB_simple_screen': bbb_simple_screen}

Per-Tool Failure Modes

ADMETlab 3.0 -- out-of-distribution prediction

Trigger: Chemistry materially unlike the service's documented training/applicability domain, such as metal-containing complexes, many peptides, PROTACs, or unusual macrocycles.

Mechanism: ADMETlab training set is drug-like organic molecules. Predictions on PROTACs, macrocycles, peptides extrapolate.

Symptom: High reported uncertainty, disagreement with neighbors or orthogonal models, or unstable conclusions under reasonable preprocessing.

Fix: Check uncertainty band; if interval is broad, do not trust point estimate. For PROTACs / macrocycles, prefer literature-derived experimental data.

hERG D-MPNN -- training data bias

Trigger: Compound is novel chemotype not in training set (drug-like but in unexplored region).

Mechanism: D-MPNN learns local chemical features; for genuinely new scaffolds, extrapolation is unreliable.

Symptom: Model predicts hERG- (false negative) for compound that experimentally inhibits.

Fix: Use ensemble + applicability-domain assessment (kNN distance, ensemble variance). If kNN distance to training set > P95, treat prediction as low-confidence.

CYP3A4 inhibitor + substrate ambiguity

Trigger: Model trained on either inhibitor OR substrate; predictions confused.

Mechanism: CYP3A4 inhibitors and substrates have similar SAR; many compounds are both.

Symptom: Both classes report > 0.5.

Fix: Two separate models (inhibitor model, substrate model); compounds that score high in both are flagged for in vitro confirmation.

SwissADME -- no API

Trigger: Wanting to batch programmatically.

Mechanism: SwissADME's terms restrict automated crawler/data-retrieval access, and no public API is documented.

Symptom: No programmatic access; manual web upload only.

Fix: Use a currently documented programmatic service and follow its access policy; for ADMETlab 3.0, verify the live API tutorial before writing a client.

PAINS as a kill filter

Trigger: Treating PAINS_A match as a categorical exclusion.

Mechanism: PAINS is calibrated against HTS assay-interference; matches do NOT predict failed drug development.

Symptom: Library purged of valid leads (curcumin analogs, polyphenol natural products).

Fix: Flag PAINS for orthogonal-assay confirmation; do not exclude pre-emptively. See substructure-search for details.

Class-imbalanced AMES dataset

Trigger: Training/predicting AMES mutagenicity.

Mechanism: Public AMES datasets can be imbalanced and differ in assay definition and curation; aggregate accuracy can therefore be misleading.

Symptom: Model reports high accuracy but predicts negative for all.

Fix: Report class balance and use suitable metrics such as PR-AUC, ROC-AUC, MCC, or balanced accuracy. Compare class weighting or resampling inside training folds without leaking validation/test data.

Reconciliation Across Models

When ADMETlab, ProTox-3.0, and an independently trained chemprop model disagree on hERG:

• All predict hERG+ -> higher concern; plan in vitro patch-clamp
• Results disagree -> inspect applicability domains, activity thresholds, and assay definitions before deciding
• All predict hERG- -> lower concern, but still consider in vitro screening for clinical candidates and novel chemotypes
• Do not count correlated models as independent evidence merely because they are hosted by different services

Common Errors

| Symptom | Cause | Fix | |---------|-------|-----| | ADMETlab API timeout | Service load, payload, or current quota | Follow live batch limits; retry with backoff and record failures | | chemprop training overfits | Random split | Use scaffold split (--split scaffold_balanced) | | hERG prediction 50/50 | Out-of-distribution | Check applicability domain | | QED calculation fails | Molecule is missing, unsanitized, or unsupported | Reject parse failures; sanitize inputs and handle calculation exceptions | | ProTox endpoints missing | Web scrape uses CSS selector | Use formal API | | BBB+ true but TPSA > 90 | Different BBB model | Use the simple physicochemical screen as an orthogonal heuristic | | Predictions inconsistent across runs | Random seed for chemprop ensemble | --seed 42 and reuse model | | Calibration mismatch | DL native probabilities not calibrated | Apply Platt scaling on validation set |

References

• Fu et al., Nucleic Acids Res. 52:W422-W431 (2024) -- ADMETlab 3.0 (DOI 10.1093/nar/gkae236).
• Shan M, Jiang C, Chen J, Qin L-P, Qin J-J, Cheng G. RSC Adv. 12:3423-3430 (2022) -- D-MPNN/MOE hERG benchmark (DOI 10.1039/D1RA07956E).
• Karim A, Lee M, Balle T, Sattar A. J. Cheminformatics 13:60 (2021) -- CardioTox-net (DOI 10.1186/s13321-021-00541-z).
• Banerjee P, Kemmler E, Dunkel M, Preissner R. Nucleic Acids Res. 52:W513-W520 (2024) -- ProTox-3.0 (DOI 10.1093/nar/gkae303).
• McShane SA et al. J. Cheminformatics 16:75 (2024) -- conformal prediction for molecular-property models (DOI 10.1186/s13321-024-00870-9).
• Heid E et al., J. Chem. Inf. Model. 64:9-17 (2024) -- Chemprop redesign (DOI 10.1021/acs.jcim.3c01250).
• Chemprop documentation, training, descriptors, and uncertainty: https://chemprop.readthedocs.io/
• ADMET-AI official repository and version notes: https://github.com/swansonk14/admet_ai
• SwissADME Terms of Use: https://www.swissadme.ch/termsofuse.php
• OECD, "Principles for the Validation, for Regulatory Purposes, of (Q)SAR Models" (agreed 2004).
• OECD, Guidance Document on the Validation of (Q)SAR Models, OECD Series on Testing and Assessment No. 69 (2007).
• OECD, "(Q)SAR Assessment Framework" (2023).
• Capuzzi et al., J. Chem. Inf. Model. 57:417 (2017) -- PAINS reality check.
• Wager et al., ACS Chem. Neurosci. 1:435 (2010) -- Pfizer CNS MPO.
• Bickerton et al., Nat. Chem. 4:90 (2012) -- QED.

Related Skills

• chemoinformatics/molecular-descriptors - Compute drug-likeness physchem
• chemoinformatics/substructure-search - PAINS / BRENK / REOS filter
• chemoinformatics/qsar-modeling - Build custom QSAR for in-house data
• chemoinformatics/molecular-standardization - Canonicalize before prediction
• machine-learning/biomarker-discovery - Adjacent ML approaches
• clinical-databases/pharmacogenomics - Patient genotype overlay